Mining performance
Recorded runtime and hashrate
Four computers logged 609.24 machine-hours on real Kryptex XELIS jobs. The two 24-thread rigs contributed 290.89 machine-hours combined. Each system used huge pages for all workers.
| Computer | Threads | Recorded hours | All sessions (kH/s) | Last session (kH/s) |
|---|---|---|---|---|
| DesktopAMD Threadripper PRO 7995WX | 192 | 171.15 | 122.698 | 123.114 |
| CPU Rig 1 | 32 | 147.20 | 12.469 | 12.462 |
| CPU Rig 2 | 24 | 143.97 | 8.561 | 8.558 |
| CPU Rig 3 | 24 | 146.92 | 9.152 | 9.152 |
“All sessions” weights each miner-displayed session average by its recorded duration. “Last session” is the average displayed when each computer ended its final beta run. The rates include user and developer-fee work. Runtime includes warmup and reconnects. Gaps between runs and starts that never reported a rate are excluded. The fleet total uses unrounded durations.
Kryptex accepted 86,134 shares, including 85,276 user shares and 858 developer-fee shares. The miners recorded 9 rejections and 195 submissions with an unresolved outcome. Sustained 140 kH/s and matched CPU package power remain unverified. Download the final runtime data ↗ · Earlier 15:20 UTC snapshot
Beta.5 development build · Kryptex trial
The Kryptex trial lasted 33 hours, 31 minutes, 18 seconds on an AMD Threadripper PRO 7995WX with 192 huge-page workers. The miner displayed a session average of 124.060 kH/s. Kryptex accepted 13,911 user shares and 137 fee shares. The miner recorded zero rejected shares and 13 unresolved submissions during connection interruptions. The CPU peaked at 83.875 °C, below the 85 °C trial guard.
The trial recorded a session-display rate. A completed-hash counter window and CPU package power were unavailable. A matched speed gain and the 140 kH/s target remain unverified.
The source preview is based on the tested development kernel and warm-handoff runtime. The public package was rebuilt with its own version metadata. The 33-hour trial predates that public rebuild.
Read the validation scope · View the sanitized trial evidence · Read the public validation notes.
Earlier beta.4 hotfix · source preview 2
New: development hotfix logged hour · 6 November 2026
60 minutes 27.49 seconds after warmup · 192 huge-page workers · 119.739756 kH/s mean of 361 rounded local displays; sampled minimum 119.517, maximum 119.978 kH/s. The measured interval accepted 444 USER shares; the full session accepted 450, with zero rejected or unresolved displayed. All native workers closed normally after a requested stop.
This is logged continuity evidence from the development hotfix, whose native library is byte-identical to the qualified public preview-2 build. The launcher retained its earlier preview-1 banner. Raw hash-counter boundaries and a monotonic interval were not saved; native drops/undefined skips, GPU activity, cooling and temperatures were not recorded for this session. This is not an exact-counter hour, a full fee cycle or a 120 kH/s guarantee.
Read the dated method, identity and limits · Inspect sanitized interval data.
Earlier short hotfix counter measurements
| Hotfix run | Completed-hash local rate | Measured interval | Accepted USER shares |
|---|---|---|---|
| First run | 119.302920 kH/s | 191.063907 seconds | 25 |
| Repeat run | 118.955001 kH/s | 190.778115 seconds | 28 |
Scope: zero rejected, unresolved or queue-dropped work was reported in these short measurements. Saved Rust answers were replayed through four C implementations for 1,031 inputs, full scratchpads were compared, and 192 actual huge-page workers were verified. Forced-zero guard tests exercised 1, 24, 32 and 192 workers. Rust was not newly executed. These checks support the observed selected-kernel repair and normal hash behavior; they do not establish long-term stability, broad hardware performance, an exact-counter hotfix hour, a fee-cycle pass or a general 120 kH/s guarantee.
The production selected kernel uses the checked path. Reference and experimental kernels retain their earlier zero-modulus behavior; the legacy void selected API deliberately traps on failure. The hotfix does not define a new consensus hash for an undefined input. Capturing and replaying a naturally failing input remains necessary to distinguish a reference edge case, an optimization discrepancy or a transient computation fault.
Inspect beta.4 hotfix evidence · Build identities and scope.
Historical, unpatched beta.4: hour and separate profiles
The records in this section belong to superseded source preview 1. Its archive and test data remain unchanged. They do not establish hotfix hour-long performance or hotfix portability.
6 November 2026 (UTC). The selected desktop build mined real Kryptex USER work. The panel records its full tested conditions; it is not a matched comparison with beta.3 or another miner.
What the hour measures
After 600 seconds of settling, the audit counted 366 rounded Local interval displays with no observed metric gaps. Their unweighted mean was 121.157224 kH/s; minimum 120.908, maximum 121.463. The conservative post-settle active USER duration was at least 3,671.545 seconds (over 61 minutes). Every counted display exceeded the guarded 120.001 kH/s threshold, allowing for its 0.001 kH/s display precision. Kryptex accepted 393 USER shares during measurement, with zero rejected or unresolved. Normal owned-process cleanup was verified.
No raw whole-hour completed count is available, so this mean is not total raw hashes divided by the hour. Settling and cleanup samples are excluded. It is not pool-side hashrate, net earnings, a statistical-confidence claim or a full developer-fee-cycle test. The source passed saved-reference/four-C/full-scratchpad qualification; Rust answers were replayed rather than newly executed.
Short desktop profiles: 24 and 32 threads
These separate Kryptex runs used the same 7995WX desktop build, huge pages, paused GPU mining and Mining100 cooling. Each followed 60 seconds of settling. Rates below use gross completed hashes divided by the measured interval; they are not the hour’s rounded-display metric or results from separate physical rigs.
| Desktop profile | Gross completed rate | Measured window and accepted work |
|---|---|---|
| 24 threads | 23.844 kH/s | 311.380 seconds · 7,424,693 completed hashes · 3 accepted USER shares. |
| 32 threads | 31.786 kH/s | 311.512 seconds · 9,901,797 completed hashes · 13 accepted USER shares. |
Three physical rigs: local builds and startup checks
These are separate computers, not the desktop thread-count profiles above. Each rig compiled the source locally, passed the full 1,031-input/four-C hash and full-scratchpad qualification, verified all requested huge-page workers, and submitted fresh Kryptex USER work. The desktop native library was not copied to them.
| Physical rig | Compiler | Observed startup result |
|---|---|---|
| Rig 1 · 32 threads | GCC 13.3.0 | 32 huge-page workers · 1 accepted USER share at first success · 0 rejected / unresolved. |
| Rig 2 · 24 threads | GCC 13.3.0 | 24 huge-page workers · 1 accepted USER share at first success · 0 rejected / unresolved. |
| Rig 3 · 24 threads | GCC 11.4.0 | 24 huge-page workers · 1 accepted USER share at first success · 0 rejected / unresolved. |
Counter boundaries for the short desktop runs are sampled across workers. The reported sensitivity ranges are 23.844299–23.844609 kH/s for 24 threads and 31.786037–31.786451 kH/s for 32 threads; these are boundary-sensitivity bounds, not confidence intervals.
Inspect the beta.4 evidence JSON · Build identities and validation scope.
Preserved beta.3 versus beta.2: matched local test
1 November 2026 (UTC). Both builds used the same GCC 14.3 compiler, normal non-PGO build mode, 192 CPU threads, ordered CPU placement and hugetlb memory. Each measurement used a fresh process loading one native library, with a 15-second warmup and a 60-second sample. Four samples per build ran in alternating order with the competing CPU miner stopped.
| Build | Weighted local rate | Scope |
|---|---|---|
| OverXelis beta.3Native library reproduced by public source preview 1 | 111.168 kH/s | All four candidate windows exceeded the 110.615 kH/s target; range 110.877–111.611 kH/s. The final public build reproduced the measured native library byte for byte. |
| OverXelis beta.2Public source preview 2, rebuilt with the same GCC 14.3 | 110.234 kH/s | Same thread placement, requested memory mode and sample durations as beta.3. |
Aggregate improvement: +0.85% (0.847585957% from unrounded totals). The median of the four paired differences was +1.022%; it is a different summary from the aggregate improvement. Four separate windows are not one uninterrupted four-minute measurement. Clock, temperature, scheduling and order variation still limit the result; no formal statistical-confidence claim is made.
Exact public package: five-minute live pool check
The final beta.3 public source package completed a normal 300-second USER pool run. Its final application-local session average was 110.870 kH/s, including startup and job waits, with 38 accepted shares and zero reported rejected, unresolved, local stale, duplicate or invalid shares. No queue drops occurred. All 192 workers used huge pages and closed normally. The displayed rate is rounded to 0.001 kH/s; it is not a pool-side hashrate estimate or a payout measurement.
Scope: this establishes a measured improvement on one desktop and a finite successful pool check. It does not establish a matched 5% advantage over BzMiner, a full live beta.3 developer-fee cycle, forced-reconnect recovery, a long soak, or performance on other computers. Earlier beta.2 fee-cycle evidence below does not qualify those beta.3 checks.
Read the beta.3 execution evidence · Preserved beta.3 qualification and limits · Beta.3 prerelease and named downloads.
Historical observations · 29 October 2026 UTC
The following records and arithmetic retain their original beta.2-era reference. They are not recalculated with beta.3 or beta.4 rates and do not rank either release against other miners.
These saved results concern earlier private development builds. Beta.2 Preview 2 corrected documentation and attribution; that correction did not add a benchmark. Historical Preview 2 correction.
Historical recorded rates and arithmetic
The calculation is (OXM reference ÷ comparison rate − 1) × 100. OXM’s reference is 110.614949 kH/s. Values below are not adjusted for fees or connection downtime. TNN uses the 04:05 UTC snapshot on 29 October 2026; HydraX uses its completed ten-minute accepted-share trial.
| Software / build | Recorded rate | Recorded-rate difference | Test context |
|---|---|---|---|
| OverXelis MinerSelected development kernel | 110.615 kH/s | Reference | 192 threads; huge pages. Four 60-second completed-hash windows after warmup. Not live qualification of the public package. |
| BzMinerv100.46 | 106.360 kH/s | +4.00%Arithmetic only | Historical mean: 1,984 reports after warmup in a roughly 34.6-hour run. Different dates and measurement method. |
| SRBMiner-MULTIv3.7.5 | 105.867 kH/s | +4.48%Arithmetic only | Five short historical trials; ten later displays and 55 accepted-share events. Startup exclusions are disclosed below. |
| TNN-Minerv0.8.0 | 101.970 kH/s | +8.48%Arithmetic only | 40 live displays; 83 user shares accepted, zero user rejected. Snapshot at 04:05 UTC; tuning excluded. |
| cpuminer-opt-suprAVX2; banner v1.40 / folder v1.47 | 95.786 kH/s | +15.48%Arithmetic only | Eight displays and three accepted shares in a 48-second startup record. Version discrepancy remains unresolved. |
| hoo_cpu / HTNv1.4.3 | 97.018 kH/s | +14.01%Arithmetic only | Five later rolling displays in a short trial; 19 accepted USER shares. Not the newer v1.4.24. |
| HydraX Minerv3.4.5 | 19.242 kH/s | +474.87%Arithmetic only | 10 min 19 sec; 30 displays and 11 accepted shares. 4 KB pages vs OXM’s 2 MiB huge pages: settings differ. |
Follow-up: accepted work and recovered logs
Added 29 October 2026 (UTC). These observations extend the original records below; they do not replace them or establish a controlled ranking.
TNN now has a successful v0.8.0 accepted-share test
Snapshot at 04:05:19 UTC on 29 October 2026, about 14 minutes 7 seconds after launch, including tuning. The miner was running when captured. 40 live HASHRATE displays averaged 101.970 kH/s (range 99.8–104.0); the final five averaged 101.560 kH/s. Automatic tuning results are excluded.
The snapshot records 83 accepted user shares and zero user rejections, plus 86 accepted developer shares and one developer low-difficulty rejection. Different share difficulties mean these counts do not establish the effective fee. The run requested 192 threads and a 1% fee. Huge pages were reported available, not independently confirmed as used; MSR optimization was disabled. This short v0.8.0 observation is not a result for v0.9.0 or a matched OXM benchmark.
HydraX now has an accepted-share trial
The second HydraX v3.4.5 session ran from 03:34:03.958 to 03:44:22.469 UTC: about 10 minutes 19 seconds. It recorded 11 accepted shares and zero rejections. Thirty displayed rates averaged 19.242 kH/s (range 19.18–19.42); the final five averaged 19.186 kH/s. The miner requested 192 threads and reported 4 KB fallback pages.
This is enough to document working pool submissions and the displayed speed of this tested configuration. It is not enough to establish a matched performance advantage: OXM used 2 MiB huge pages, the software reports differ, and a complete fee-cycle or longer-run payout comparison was not established. The current table uses this accepted-share result. The initial zero-share trial remains in the original evidence summary and archive note below.
OXM: hashing speed is not the downtime-inclusive average
The retained desktop beta.2 session that began at 02:18:38 UTC on 29 October 2026 had an uninterrupted first interval of approximately 44 minutes 52 seconds, ending with a 110.707 kH/s session average. That supports a live running rate close to the retained 110.615 kH/s development reference. It is only about 0.08% higher, not evidence of a substantially new speed level.
The whole session lasted about 71 minutes 34 seconds, with 52 minutes 43 seconds of recorded active user work and about 18 minutes 51 seconds not active. Its final all-time average was 81.730 kH/s. An active-time-only reconstruction is approximately 110.969 kH/s, but uses rounded display values and second-resolution timestamps, not an exact stored counter boundary.
The final recovery interval included short readings up to 113.595 kH/s. Those do not replace the longer reference. The session ended with 293 accepted, zero rejected, and 30 unresolved submissions; unresolved means the outcome was not confirmed, not that the pool rejected the share. No developer-fee window occurred. Recovery reliability and unknown submissions still need attention.
Historical CPU tests recovered from saved logs
| Miner / record | Observed rate | Evidence and limitations |
|---|---|---|
| SRBMiner-MULTI v3.7.5 14 October 2026 | 105.867 kH/s Ten later displays | Five short 192-thread CPU trials with reported 2 MB huge pages and 55 accepted-share events in total, with no rejection events in those logs. Later readings span 105.43–106.17 kH/s. For this descriptive mean, the first of three displays in each trial is excluded; all 15 readings are retained in the evidence file. All 15 average 103.512 kH/s. Relative to the OXM reference, the later-display arithmetic difference is +4.48%, not a matched speed claim. |
| cpuminer-opt-supr 15 October 2026 | 95.786 kH/s Eight displays | 192-thread AVX2 run, huge pages reported, three accepted shares and zero recorded rejections. A very short 48-second record includes startup; displays span 94.95–98.55 kH/s. Arithmetic difference from the OXM reference: +15.48%. The directory is labeled v1.47 but the startup banner says v1.40, so the version discrepancy remains explicit. |
| TNN-Miner v0.9.0 15 October 2026 | No usable rate | The retained run requested XELIS v3 with 192 threads, then logged free(): invalid pointer during tuning. It produced no usable hashrate in that log. A failed startup is not a 0 kH/s benchmark. A successful v0.8.0 retest is now reported above; it does not change this older failure record. |
Older SRBMiner records varied substantially: four July 16–17 v3.4.6 sessions averaged about 99.961–102.182 kH/s across all their displays, while a July 14 session averaged 51.295 kH/s. A truncated October 13 v3.7.5 excerpt averaged 77.430 kH/s. We retain this context rather than selecting only the slower records for a promotional comparison.
Download the follow-up observation summary for all selected readings, counts, source hashes, exclusions, and limitations. Raw personal logs and private machine paths are not published. Historical logs used here were checked against the saved collection index; newer observations have their own recorded log identities.
Earlier HydraX trial — preserved, not the current comparison
The initial 87-second v3.4.5 trial averaged 18.950 kH/s across four displays, with 4 KB fallback pages and no accepted share before stopping. Its +483.72% arithmetic gap is historical only. The main table now uses the longer run with 11 accepted shares.
Precise inputs, consistent calculations
Percentages use the unrounded rates in the evidence files and are rounded to two decimal places. The table shows rates to three decimals. Rounded recollections such as 96 or 18 kH/s are not calculation inputs. The longer accepted-share HydraX observation replaces the original short trial in the current table.
We report miner-by-miner values rather than one speed range across all six projects. The available evidence supports specific, qualified observations—not a blanket ranking. The follow-up above adds historical SRBMiner-MULTI and cpuminer-opt-supr observations. TNN-Miner now has a successful v0.8.0 snapshot; the earlier v0.9.0 tuning failure remains in the historical record. None is a matched OXM comparison.
Version and memory context
The hoo_cpu trial used v1.4.3, while its official download page listed v1.4.24 when checked. A separate attempt with the newer release did not produce a qualified hashrate result in the reviewed records. The older observation is not presented as the current release’s performance. HydraX’s 4 KB memory mode differs from OXM’s huge-page configuration. That difference must not be hidden in a headline.
Saved rolling displays for hoo_cpu, in order: 82.50, 97.48, 97.17, 96.98, 96.90, 96.56 kH/s. Initial, archived HydraX displays: 18.96, 18.95, 18.95, 18.94 kH/s. See the sanitized machine-readable observation summary for exact values, dates, and source-record hashes. It omits payout addresses, private paths, and raw personal logs; it is a summary, not a substitute for independently reproduced runs.
Download the preserved 29 October calculations and TNN snapshot ↓
What a fair comparison needs
Record the CPU and operating system, exact miner versions and binary identities, compiler/build flags, requested and actual thread placement, actual memory mode, clock and power settings, and temperatures. Use the same work source and pool where practical. Run miners sequentially without competing CPU workloads, and balance run order to reduce heating and order effects.
Choose warmup and measurement durations before testing. Keep all planned runs, including failures, and repeat each configuration. Distinguish local completed-hash rates from pool estimates and report accepted, rejected, stale, and unresolved work. Compare accepted difficulty-weighted work over sufficient time; share counts alone are not comparable when difficulties differ.
For earnings or fee-efficiency claims, cover complete fee cycles and include reconnection overhead and user-only accepted work. Publish per-run results, variation, exclusions, and limitations. Never convert a startup peak, a crash, or a zero-share trial into proof of a competitor’s maximum speed.
Official projects and published fees
Documentation was checked on 29 October 2026 (UTC). These are publisher statements, not an audit of each miner’s effective fee. Pool fees are separate. Always check the exact version you download.
| Project | CPU support | Published miner fee | Evidence / official source |
|---|---|---|---|
| BzMiner (official project, opens in a new tab) | CPU-only XELIS operation is documented. Algorithm: -a xelis. | 1% | 106.360 kH/s observedv100.46 · See test scope and limitations. Official documentation for BzMiner (opens in a new tab) |
| SRBMiner-MULTI (official project, opens in a new tab) | The supported-algorithm table lists xelishashv3 with CPU support. | 1.5% | 105.867 kH/s observedv3.7.5 · See test scope and limitations. Official documentation for SRBMiner-MULTI (opens in a new tab) |
| TNN-Miner (official project, opens in a new tab) | Open-source XELIS v3 support; --xel-v3 and --no-gpu select CPU-only mining. | 2.5% default; configurable to 1% | 101.970 kH/s observedv0.8.0 · See test scope and limitations. Official documentation for TNN-Miner (opens in a new tab) |
| cpuminer-opt-supr (official project, opens in a new tab) | Open-source CPU miner listing xelishashv3. The specific -supr fork matters. | 0% built-in miner fee | 95.786 kH/s observedAVX2; banner v1.40 / folder v1.47 · See test scope and limitations. Official documentation for cpuminer-opt-supr (opens in a new tab) |
| hoo_cpu — HTN CPU Miner (official project, opens in a new tab) | XELIS v3 on x86-64 CPUs with AVX2. Linux binary; Windows use through WSL. | 1.25% for XELIS v3 | 97.018 kH/s observedv1.4.3 · See test scope and limitations. Official documentation for hoo_cpu — HTN CPU Miner (opens in a new tab) |
| HydraX Miner (official project, opens in a new tab) | Documents XelisHash v3 CPU mining and Windows/Linux x64 support. | 1% | 19.242 kH/s observedv3.4.5 · Memory settings differ. Official documentation for HydraX Miner (opens in a new tab) |
Built with appreciation
Thank you to Slixe (opens in a new tab) for creating XELIS and for his development work, and to all XELIS contributors. We deeply value the project’s work on privacy, open participation, and programmable digital money. Learn about XELIS (opens in a new tab) from its official project site.
OXM remains independent: no affiliation with or endorsement by Slixe, the XELIS project, Kryptex, or the miner developers listed here is implied. Third-party source acknowledgments and licenses remain in the download.